Covariance function

The covariance function, or kernel, describes the spatial covariance of a random field.

stats++

In stats++, a covariance function is defined by implementing a derived class of the CovarianceFunction class:

#include <memory>
#include <vector>
#include "jScience/linalg/Vector.hpp"
 
class CovarianceFunction
{
 
public:
 
    CovarianceFunction(const std::vector<double> &th);
    ~CovarianceFunction();
 
    virtual std::unique_ptr<CovarianceFunction> clone() const = 0;
 
    virtual double              cov(const Vector<double> &x1, const Vector<double> &x2) = 0;
    virtual std::vector<double> ddtheta(const Vector<double> &x1, const Vector<double> &x2);
    virtual Vector<double>      ddx(const Vector<double> &x, const Vector<double> &x2);
 
protected:
 
    std::vector<double> theta;
 
};

This class is purely abstract, as the clone() and cov() functions must be defined in the derived class; note that implementations of ddtheta() and ddx() are not required, unless certain functionality is desired (e.g., optimization of hyperparameters).

Example

An example for a simplified squared exponential covariance function:

#include <memory>
#include <vector>
#include "jScience/linalg/Vector.hpp"
 
class SquaredExponential : public CovarianceFunction
{
 
public:
 
    SquaredExponential(const std::vector<double> &th) : CovarianceFunction(th) {};
    ~SquaredExponential() {};
 
    std::unique_ptr<CovarianceFunction> SquaredExponential::clone() const
    {
        return std::unique_ptr<CovarianceFunction>(new SquaredExponential(*this));
    }
 
    double cov(const Vector<double> &x1, const Vector<double> &x2)
    {
        Vector<double> x1mx2 = x1 - x2;
 
        return ( this->theta[0]*this->theta[0]*std::exp(-dot_product(x1mx2, x1mx2)/(2.0*this->theta[1]*this->theta[1])) );
    }
 
};